funes Gives Coding Agents Persistent, Local Memory
Hugging Face has introduced funes, an open-source memory layer for coding agents including Claude Code, Codex, pi and Hermes. The tool turns existing agent session traces into a searchable local memory, preserving decisions, failed approaches and their rationale instead of leaving them in separate logs. A single command creates an index, adds recall and retrieval tools, and updates the index after completed turns.
funes combines vector and BM25 search, ranking fusion, cross-encoder reranking and recency weighting. Results contain the original text and provenance, including the agent, timestamp, session and turn, while a get command opens surrounding context. Embeddings and reranking run locally, and no account or hosted model is required for the default setup.
Users can bind a memory to a private-by-default Hugging Face dataset so it can be shared across machines or with a team. Credentials are redacted before publication, with an additional scan for possible secrets. The article reports that, in a handoff-versus-recall benchmark, recall was four to eight times cheaper than a written handoff and preserved passages that compaction could lose.
Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.